3 ms·

It's economics. I can attest to the fact that most five-year-old chips are good enough for 95% of software I want to use. So there's no real demand outside of b

by api 14y ago

It's economics. I can attest to the fact that most five-year-old chips are good enough for 95% of software I want to use. So there's no real demand outside of big data and data center applications. The only way to keep pushing performance would be for software writers to invent new and compelling things to do that genuinely require more power. It can be something as simple as a game-- I was a teenager when Quake came out, and I remember scraping together as much money as I could to get the absolute fastest machine I could (A P120) to play it. Before Quake came out, my old 486 DX2/66 was fast enough for the Linux hacking, adventure games, and BBS/Internet stuff I was doing at the time. We need another Quake... or something equally cool. We need killer apps. Otherwise computers are going to go over like the aerospace industry did. We stopped pushing the envelope there in the 70s, again for economic reasons.

EthanHeilman 14y ago
Technologies like the oculus rift will force a resolution and latency arms race (the resolution arms race is already in full swing thanks to attempts to surpass apple's "retina display"). Of course this arms race will take place mostly on the GPU and display side of things. I see the CPU becoming more of a coordinator between custom processors (think the video decoders on iphones) than the main show. * What are the computational bottlenecks for driverless cars? Vector operations? * If every car needs a high performance computer to drive how will this change the semi-conductor landscape?
qznc 14y ago
The point is, it must be about something noticeable other than the graphics. For example, a good AI should easily be able to exhaust all the cores you can throw at it. However, it is much harder to sell. In those days they could show you some pictures of Quake and you know you wanted that. It is harder to make "pictures" of AI behavior. Alternatively, exchange AI for physics simulation. Maybe the sandbox genre (Minecraft, Dwarf Fortress) will grow. Simulating huge worlds requires lots of processing power.
- api 14y ago
  I did a lot of work on evolutionary computation a while back. Now there's something that can soak up some cycles. It's also embarrassingly parallel. (For those who want to trot out the "GAs don't work well," "evolution is slow," etc. mantras, please delve into the literature first. EC has some interesting and unique results, the most important of which is that it's the only "AI" technique I'm aware of that can demonstrate true creativity. EC can invent novel structures ex-nihilo. Everything else can just optimize, find maxima, etc. Another way of saying it is that EC can find maxima involving unknown, novel parameters that it invents along the way, while everything else requires that you know all the parameters to be optimized ahead of time. So EC can work with unknown unknowns, not just unknown values of knowns.)

by api 14y ago

EthanHeilman 14y ago

Technologies like the oculus rift will force a resolution and latency arms race (the resolution arms race is already in full swing thanks to attempts to surpass apple's "retina display"). Of course this arms race will take place mostly on the GPU and display side of things. I see the CPU becoming more of a coordinator between custom processors (think the video decoders on iphones) than the main show. * What are the computational bottlenecks for driverless cars? Vector operations? * If every car needs a high performance computer to drive how will this change the semi-conductor landscape?

qznc 14y ago

The point is, it must be about something noticeable other than the graphics. For example, a good AI should easily be able to exhaust all the cores you can throw at it. However, it is much harder to sell. In those days they could show you some pictures of Quake and you know you wanted that. It is harder to make "pictures" of AI behavior. Alternatively, exchange AI for physics simulation. Maybe the sandbox genre (Minecraft, Dwarf Fortress) will grow. Simulating huge worlds requires lots of processing power.

api 14y ago
I did a lot of work on evolutionary computation a while back. Now there's something that can soak up some cycles. It's also embarrassingly parallel. (For those who want to trot out the "GAs don't work well," "evolution is slow," etc. mantras, please delve into the literature first. EC has some interesting and unique results, the most important of which is that it's the only "AI" technique I'm aware of that can demonstrate true creativity. EC can invent novel structures ex-nihilo. Everything else can just optimize, find maxima, etc. Another way of saying it is that EC can find maxima involving unknown, novel parameters that it invents along the way, while everything else requires that you know all the parameters to be optimized ahead of time. So EC can work with unknown unknowns, not just unknown values of knowns.)